Failure analysis on a water pump based on a low-cost MEMS accelerometer and machine learning classifiers

Abstract

This work presents a failure diagnosis tool for a water pump using a low-cost MEMS accelerometer. It was inserted three types of failures: rotor blade (new and damaged), pump soleplate tightness (stiff or loose), and cavitation, in this case on three conditions: none, incipient and severe, totaling twelve fault combinations. These conditions were tested under two different speeds to perform the diagnosis, totaling twenty-four tests. In all cases, the vibration signals from axes X, Y, and Z were acquired. Some features extracted from the vibration spectra from X-axis were used to compose the dataset. These data were analyzed employing logistic regression, a linear support vector machine (SVM), and an artificial neural network multilayer perceptron (ANN-MLP). We compared these three techniques of machine learning and evaluated which one was able to obtain the most accurate result. Using the ANN-MLP, the system was able to detect all three types of failures inserted, with about 100% of accuracy on the rotor blade condition, 92% for anchorage faults, and about 99% accuracy on cavitation state. As a conclusion, it is demonstrated that this classifier algorithm can be used to process the data from the low-cost MEMS accelerometer in predictive maintenance as an accurate tool.